Fixed price or time and materials for AI projects

Fixed price works when the scope is genuinely known - a defined integration, a specified document type, an agreed set of intents. Time and materials works when the work is genuinely exploratory. The expensive mistake is fixed price on an unknown scope, because the vendor prices their risk into your invoice and then defends the boundary instead of solving your problem.

AI projects contain both kinds of work, usually in sequence. That is why the contract shape matters more here than on ordinary software.

Key takeaways

Fixed price does not transfer risk. It prices it, and you pay the premium whether or not the risk occurs.
The parts of an AI build that are genuinely unknowable are usually data quality and permissions.
A fixed price quoted before anyone has looked at your data is padded, guessed, or will be revised.
Phase the contract: fixed fee to find out, then fixed price on what is now known.
Watch for the change-request incentive. It is the main failure mode of fixed price.

Who this applies to

You are contracting an AI build between roughly $15,000 and $150,000 and choosing a commercial structure. Applies equally to agencies and contractors.

What each structure actually does

Neither structure makes a project cheaper. Both allocate risk, and someone pays for it either way.

	Fixed price	Time and materials
Who carries scope risk	Vendor	You
What you pay for uncertainty	A premium built into the price	Only the hours actually used
Behaviour when reality differs	Change requests, boundary defence	Continue, re-plan
Your budget certainty	High	Low without a cap
Vendor incentive	Finish fast, minimise scope	Be thorough, no natural stop
Best when	Scope is known	Scope is genuinely unknown

The fixed-price premium is typically 15-40% on work with real uncertainty. That is not vendors being greedy; it is what carrying the risk costs. If the risk does not materialise, you paid for it anyway. If it does, you were protected and got good value. It is insurance, and it should be evaluated as insurance.

Which parts of an AI build are knowable

This is the practical question, because most AI projects contain both kinds of work.

Reasonably knowable, so fixed-priceable:

Integrating with a documented API you have access to
Building an evaluation harness and scoring pipeline
A defined set of intents or one specified document type
Deployment, monitoring and alerting setup
Training and handover

Genuinely unknowable until someone looks:

The state of your data. Missing fields, duplicates, legacy status values, encoding surprises.
Permission modelling, if your access rules live in people’s heads rather than a system.
How much of your documentation is true.
Whether the accuracy the business case needs is achievable at all on your data.

That last one is the important one. Nobody can fix-price an accuracy target on data they have not seen. A vendor who guarantees 90% resolution before reading your tickets is either padding heavily or will renegotiate.

Nobody can fix-price an accuracy target on data they have not seen. The diagnose phase exists to empty the right-hand column, after which a fixed price is honest rather than padded.

The structure that works

The pattern that avoids the worst of both:

Fixed fee for a short diagnosis. Small, defined, days not weeks. The vendor reads your data, your systems and your documentation. Output: a scope, a baseline measurement, and a costed roadmap you keep regardless of what happens next.
Fixed price for the now-known build. After step one the unknowns have collapsed, so a fixed price is fair rather than padded and the vendor is not pricing fear.
Time and materials, capped, for the genuinely exploratory remainder if any exists. Data remediation is the usual candidate.
A retainer for operation. Ongoing work is not a project and should not be contracted as one.

This is how we structure engagements - a two-day diagnosis at $2,500, then a fixed price for a defined scope. The commercial logic is straightforward: we would rather charge a small amount to remove uncertainty than a large amount to absorb it, and you get a roadmap that is useful even if you take it elsewhere.

The failure mode of fixed price

Fixed price creates one specific bad incentive worth naming plainly: once the price is set, every discovery becomes a negotiation.

You find mid-project that the agent also needs to handle refunds. Under fixed price that is a change request, a re-quote, and a delay. Under time and materials it is Tuesday.

Worse is the quiet version. A fixed-price vendor under margin pressure does not usually argue with you. They deliver exactly what the specification says, including the parts they can see are wrong, because fixing them is unpaid. You get a system that matches the document and misses the point.

Two protections. First, make the specification outcome-shaped rather than task-shaped - “resolves these 40 intents at 65% on a held-out set” rather than “builds a chatbot with these features”. Second, include a small change allowance in the contract, perhaps 10%, that either side can draw on without renegotiating. It removes the friction from small honest discoveries.

The failure mode of time and materials

The mirror problem: no natural stopping point.

Exploratory work expands to fill available budget, and AI work is unusually good at generating plausible next steps. There is always another retrieval strategy to try.

Protections: a hard cap with a mandatory re-approval, a weekly written progress note against the target metric rather than against hours, and a defined exit criterion agreed before starting. “We stop when the eval set scores 65% or when we have spent $20,000, whichever comes first” is a good sentence to have in a contract.

What to actually ask for

For a typical AI build:

Diagnosis: fixed fee, small, roadmap yours to keep.
Integration and build against known systems: fixed price.
Data remediation: time and materials with a cap, or excluded and handled by you.
Accuracy target: never contractually guaranteed before a baseline exists. After a baseline, a target with a defined remedy is reasonable.
Operation: monthly retainer with named deliverables.

If a vendor pushes fixed price on all of it before any discovery, ask how they priced the data risk. The answer tells you whether they have done this before.

Why we quote it in phases

It costs us deals, so it is worth explaining rather than asserting.

A competitor quoting a single fixed price for the whole build looks simpler and often looks cheaper, because their number does not yet include the things they have not found. Ours arrives in two parts and the second part is not knowable on day one. In a procurement comparison that reads as less certainty, and sometimes we lose on it.

We hold the line because the alternative is worse for both sides. A fixed price on unseen data is either padded enough to cover the bad case, in which case most clients overpay, or it is not, in which case the project reaches the point where the vendor must either absorb a loss or start defending scope. Neither produces a system anyone is happy with.

The version we can defend: charge a small fixed fee to remove the uncertainty, hand over a roadmap that stands on its own, then quote firmly on what is now actually known. If you take the roadmap to another vendor, that is a fair outcome and it has happened.

When neither structure saves you

When the requirement is not agreed internally. No contract shape survives a client who has not decided what they want. Settle the target metric before the commercial conversation.

When there is no internal owner. A perfectly contracted build with nobody to receive it produces an orphan system.

When the budget covers only the build. If running cost has no budget line, the contract structure is irrelevant. A system you cannot afford to operate is a liability you paid for.

Frequently asked questions

Is fixed price safer for a first AI project?

Usually yes for the build phase, provided a diagnosis happened first. The budget certainty is worth the premium when the organisation has no prior experience to calibrate against.

What is a reasonable fixed-price premium?

15-40% over an honest time estimate, depending on how much genuine uncertainty remains. Above that, the vendor is pricing risk they should have removed with discovery.

Should we ask for a fixed price with a penalty clause on accuracy?

Only after a baseline exists. Penalty clauses on unmeasured targets produce defensive scoping and conservative systems that escalate everything, which is technically compliant and practically useless.

Can we do time and materials with a not-to-exceed cap?

Yes, and it is a good middle ground. Understand that a vendor will manage toward the cap as a budget, so pair it with an outcome criterion rather than only a spend ceiling.

How do we compare a fixed-price quote against a T&M quote?

Estimate the T&M total at the vendor’s stated rate and team size, add your own contingency, and compare that against the fixed price. The difference is the insurance premium. Then decide whether you want the insurance.

Next step

The two-day diagnosis exists to collapse the unknowns before anyone quotes a build - and the costed roadmap it produces is yours regardless of who does the work.